S. Hochreiter and J. Schmidhuber, "Long Short-Term Memory," in Neural Computation, vol. 9, no. 8, pp. 1735-1780, 15 Nov. 1997, doi: 10.1162/neco.1997.9.8.1735.

[zhangyuyi99]

S. Hochreiter and J. Schmidhuber, "Long Short-Term Memory," in Neural Computation, vol. 9, no. 8, pp. 1735-1780, 15 Nov. 1997, doi: 10.1162/neco.1997.9.8.1735. https://www.connectedpapers.com/main/44d2abe2175df8153f465f6c39b68b76a0d40ab9/Long-Short%20Term-Memory/graph

Summary

RNN cannot deal with long term memory. The proposed LSTM solved this problem with the multiplicative gate units controlling the propagation of error flows. LSTM solves complex, artificial long-time-lag tasks.

Error flow is truncated once it “wants” to leave memory cells or gate units. Therefore, no connection shown above serves to propagate error back to the unit from which the connection originates (except for connections to output units), although the connections themselves are modifiable,

Thoughts

1. Is piano playing a long time task? With in what time interval will the trajectory planning of previous step affecting the next step? If the player need to consider the future notes to press, does it make more sense to use a bidirectional LSTM? Can we make LSTM forget more frequently, to reduce the time cost and at the same time maintain the control performance?

2. For single finger pressing, will RNN be enough for controlling?

3. Has LSTM been used for designing control system? How is the performance?

4. By changing the number of memory cell blocks, we can decide how much the network will memorize.

[zhangyuyi99]

For the LSTM control system, the control sequence is defined once the midi sequence is specified. Will the control algorithm be robust to changing environment? What if there are disturbs or delays or mistakes during the playing process?